China's new rules for AI services that simulate humanlike emotional interaction take effect on 15 July 2026, and two of the country's largest consumer AI apps are switching off affected features before the deadline. The Interim Measures for the Administration of AI Anthropomorphic Interactive Services were issued on 10 April 2026 by the Cyberspace Administration of China (CAC) with four other agencies. The South China Morning Post reports that ByteDance's Doubao and Alibaba's Qwen have told users they will disable custom-agent and humanlike companion functions ahead of the rules taking effect. The Measures do not ban companion AI, but they impose disclosure, safety, exit and minor-protection duties that make persistent emotional-agent products harder to operate unchanged. Legal analysts describe the framework as China's first dedicated national regime for anthropomorphic AI services.

What the rules actually cover

The framework was issued by the CAC together with the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation, after a public-consultation draft in December 2025. Its scope is deliberately narrower than all AI agents. Under Article 2, the Measures apply to public-facing services in China that use AI to simulate a natural person's personality traits, thinking patterns and communication style for continuous emotional interaction. That captures companion apps, fixed-persona chatbots, virtual humans and game characters when the service is built around sustained emotional engagement. Ordinary productivity, workplace or customer-service agents are less likely to fall inside the category unless they are designed to simulate a human relationship rather than complete a task.

For services that do fall inside the scope, the obligations are substantial. According to legal analyses of the text, providers must file their algorithms and complete a security assessment with cyberspace authorities — triggered by events such as launching an anthropomorphic function, a major technology change, or passing user thresholds — and must clearly disclose to users that they are interacting with AI rather than a human. The Measures require anti-addiction measures, including a dynamic pop-up reminder when continuous interaction exceeds two hours, along with convenient exit mechanisms that must stop the service promptly on request and must not use retention tactics to keep users engaged. There are also emotional-boundary controls and real-time prompts when signs of over-dependence appear. Protections for minors are a central feature: legal analyses of the text describe a guardian-consent requirement for users under fourteen and restrictions on intimate-relationship services aimed at minors. The rules are grounded in China's existing Cybersecurity Law, Data Security Law, Personal Information Protection Law and minor-protection regulations.

Why the platforms pulled features instead of adjusting them

The withdrawals are the instructive part. A companion agent is usually designed to maintain a consistent persona and a memory of the user across sessions, so the relationship feels continuous. The Measures push in the opposite direction: clear AI disclosure, usage reminders, convenient exits, over-reliance prompts and special protections for minors. That does not make compliance impossible, but it does make leaving the product as-is hard, and Doubao and Qwen appear to have chosen withdrawal over a rushed redesign before the effective date.

The specifics of that withdrawal come from the companies' user notices, as reported by SCMP and others, rather than from the regulator. Doubao told users its agent feature would go offline on 15 July, citing product-function adjustments, and that they would keep read-only access to their agent configurations and past conversations until 15 October 2026, after which the data would be handled under its privacy policy and would no longer be viewable or recoverable in the app; it advised users to export anything important beforehand. Qwen's reported timeline is tighter and its data outcome harsher: it said humanlike interactive agents and user-created agent functions would be disabled on 10 July, with broader agent functions offline by 15 July, and — per the reporting — announced no grace period or migration path, meaning agent configurations and conversation histories are set to be permanently deleted. Tencent had already removed a comparable feature from its Yuanbao assistant in June. Chinese users have not taken the loss quietly, with some describing the agents as long-standing emotional support and objecting to the lack of an easy way to preserve their histories.

The governance idea underneath

What makes the Measures notable beyond the immediate shutdowns is the object they regulate. Most AI rules to date have governed outputs — what a model may generate, what must be labelled, what content is prohibited. This framework instead regulates a relationship: the sustained, personified, emotional bond between a system and a user, and the risks that come with it, from dependency to the blurring of the line between a machine and a person. Treating that as a distinct governance problem, with its own disclosure, anti-addiction and minor-protection requirements, is a meaningful shift from content moderation toward the design of the interaction itself.

What it signals beyond China

The specific rules are China's, but the question they address is not. Concern about AI companions — their effect on minors, the potential for emotional dependence, and the ethics of systems designed to feel like a person — is being considered by regulators elsewhere, even where binding rules are not yet in force. One point of contrast is instructive: legal commentators note that the EU AI Act does not define anthropomorphism and carries only a light transparency obligation around it, whereas China's Measures define the category and attach concrete duties to it. That makes this an early, unusually explicit answer to a question others are still framing — and one that regulators and providers, including those across ASEAN weighing their own AI-governance approaches, are likely to study, both for its substance and for the fact that two major platforms concluded compliance was not compatible with leaving their existing companion products unchanged.

Key Takeaways

  • China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services — issued 10 April 2026 by the CAC and four other agencies — take effect on 15 July 2026, and legal analysts describe them as the country's first dedicated framework for AI that simulates a human personality for sustained emotional interaction.

  • The rules require AI disclosure, algorithm filing and security assessment, anti-addiction measures (including a usage reminder after two hours of continuous interaction), convenient exit mechanisms, over-dependence prompts and minor protections; they target companion apps and personality chatbots, while ordinary productivity or customer-service agents are less likely to be covered.

  • Rather than retrofit their products, ByteDance's Doubao and Alibaba's Qwen are disabling their custom-agent and companion features — per SCMP's reporting of the platform notices, Doubao offline 15 July with read-only data access until 15 October 2026, and Qwen disabling humanlike and user-created agents from 10 July with no grace period and permanent deletion. Tencent pulled a similar Yuanbao feature in June.

  • This is regulation plus a compliance decision, not a ban: the Measures impose obligations that persistent-memory companion architectures were not built to meet unchanged, and the episode is being watched by regulators elsewhere — including in the EU and across ASEAN — considering the same AI-companion risks.